MiniMax’s M3.1 Flash Cuts Token Prices. Thinking Settings Shape the Bill

MiniMax released M3.1 Flash on Monday with a price low enough to invite a fresh round of experiments: 15 cents for a million input tokens and 60 cents for a million output tokens. Developers pay to send material to the model and receive its responses. Tokens are the pieces of text and other content that the service counts for billing. MiniMax announcement

The company says the model cuts overall agent-task costs by 70% from its predecessor and runs nearly three times faster. It reports 73.7% on DeepSWE v1.1, a software-engineering benchmark, and 82% on CyberGym, a vulnerability-reproduction test. These are MiniMax's reported results; SnowRock has not reproduced the runs.

For teams building agents, programs that let a model operate software and call tools, the most revealing details sit in MiniMax's pricing and compatibility documents. They describe several ways an inexpensive call can become a much larger bill, or a useful task can become substantially cheaper.

A lower rate, with more than one comparison

MiniMax's published price table lists M3.1 Flash at $0.15 for input, $0.60 for output and $0.015 for cached input, per million tokens. A cache lets the service reuse previously processed material at a lower rate when the request qualifies. Official pricing

The same table lists standard M3 prices of $0.30 for input and $1.20 for output when input stays at or below 512,000 tokens. Against those rates, Flash halves the price for an identical amount of uncached text. M3's longer-input tier costs more. The companywide claim of 70% lower agent-task costs therefore describes a different comparison from a simple 50% reduction in these two listed unit prices.

For an illustrative workload with one million uncached input tokens, spread across requests within M3's lower input tier, and 100,000 output tokens, Flash's model charge would be 21 cents: 15 cents plus six cents. The same quantities at M3's lower standard tier would cost 42 cents. These calculations exclude tools, infrastructure, failed attempts and human review. They describe a billable quantity, not an observed customer job.

The default thinking setting matters

MiniMax's Anthropic-compatible API, a connection format that lets developers reuse familiar client software, supports a million-token context window for Flash. More consequentially, its documentation says the model always uses thinking and defaults to maximum effort when no effort level is specified. Developers can select lower levels. Compatibility documentation

Thinking generates additional model output while the system works toward an answer. MiniMax says higher effort produces more thinking tokens and longer waits. Its output limit includes those tokens. A limit set too low can therefore end a response before a usable answer appears.

That creates a practical migration task. A developer changing only the model name should inspect the request settings and returned usage before assuming that a cheaper published rate produces the expected saving. MiniMax also lists some familiar parameters, including temperature, as ignored by this interface. Existing software can send a setting without changing the model's behavior.

This is documented interface behavior, not evidence that every application will incur extra cost. It explains why a reproducible price comparison needs the actual request, thinking level, usage and successful outcome.

The benchmarks ask different questions

DeepSWE's creators designed 113 original engineering tasks across 91 open-source repositories and five languages. Rather than collecting fixes already merged into public repositories, they wrote new tasks and verifiers to assess requested functionality. Their paper addresses a problem in software testing: a model can appear capable if it has already encountered the answer or if a test accepts an incomplete repair. DeepSWE research

That makes the benchmark relevant to a coding model, while leaving a company's own codebase, dependencies and review rules to be tested separately. MiniMax's launch identifies version 1.1; a comparison should retain that version rather than mix scores from different releases.

CyberGym asks a more specific security question. Its Berkeley-led team assembled 1,507 historical vulnerabilities from 188 software projects. Agents receive a vulnerable codebase and a description, then try to produce an input demonstrating the bug. A successful reproduction must trigger the vulnerable version and fail to trigger the patched version. CyberGym methodology

That is useful evidence about finding a reproducible failure. A complete defensive assignment can also require discovering the problem without a description, writing a repair and preserving other functionality. CyberGym's researchers maintain separate evaluations for different stages of that process.

Finishing a workflow requires another measure

BankerToolBench, another benchmark named in MiniMax's release, was developed with 502 investment bankers. It asks agents to navigate data rooms and financial tools and create deliverables such as spreadsheets and presentations. The researchers say some tasks take bankers up to 21 hours. BankerToolBench paper

The authors' premise is that existing tests lack enough fidelity to assess “economically meaningful progress on professional workflows.” That distinction helps explain demand for more comprehensive tests. Producing a plausible paragraph differs from delivering a spreadsheet and presentation that remain consistent with source documents and instructions.

Computer-use testing adds its own comparability problem. OSWorld 2.0's authors built 108 long workflows, with a median human completion time of roughly 1.6 hours. Their work distinguishes full completion from partial progress. Its repository warns researchers to keep task files, websites and software versions aligned. A score can change meaning when the environment changes. OSWorld 2.0 research, release instructions

Research organization METR has explained another distinction in its developer-productivity research: autonomous benchmark runs and human-assisted software development measure different relationships between a person and a model. A programmer can rescue an otherwise failed attempt, then spend time reviewing the correction. That time belongs in a business comparison even when it never appears in the model provider's token bill. METR's methodology discussion

For a buyer assessing Flash, these sources support a concrete comparison: record the full model and tool bill for each attempted task, then divide by tasks that pass the required checks. Keep human correction time alongside that figure. This is an evaluation method, not a performance result SnowRock has measured.

MiniMax's launch says downloadable weights will come later. It does not give a release date in the announcement reviewed for this article. Until that changes, developers can examine the available service, preserve their request settings and test whether its low rates reduce the cost of work they can actually accept.